Pavement Management With Dynamic Traffic and ANN: A Case Study of Montreal

Page 1 of 35
For personal use only. This Just-IN manuscript is the accepted manuscript prior to copy editing and page composition. It may differ from the final official version of record.
Pavement Management with dynamic traffic and ANN: a case study of
Montreal
Can. J. Civ. Eng. Downloaded from www.nrcresearchpress.com by San Diego (UCSD) on 12/14/15
Md. Shohel Reza Amin
Research Assistant, Dept. of Building, Civil and Environmental Engineering,
Concordia University, 1515 St. Catherine Ouest
Montreal, QC H3G1M8, Canada
Tel: +14389364119. Fax: +1-514-848-7965.
Email: md_amin@encs.concordia.ca
(Corresponding Author)
Luis E. Amador-Jiménez
Associate Professor, Dept. of Building, Civil and Environmental Engineering,
Concordia University, 1515 St. Catherine Ouest
Montreal, QC H3G1M8, Canada
Tel: +1-514-848-2424 (ext. 5783). Fax: +1-514-848-7965.
Email: luis.amador@concordia.ca
1
For personal use only. This Just-IN manuscript is the accepted manuscript prior to copy editing and page composition. It may differ from the final official version of record. Page 2 of 35
Pavement Management with dynamic traffic and ANN: a case study of
Montreal
Abstract
This study improves pavement management system developing a linear programming
optimization for the road network of the city of Montreal with simulated traffic for a period of 50
years and deals with the uncertainty of pavement performance modeling. Travel demand models
are applied to simulate annual average daily traffic (AADT) every 5 years. A Backpropagation
Neural Network (BPN) with a Generalized Delta Rule (GDR) learning algorithm is applied to
develop pavement performance models without uncertainties. Linear programming of lifecycle
optimization is applied to develop maintenance and rehabilitation strategies to ensure the
achievement of good levels of pavement condition subject to a given maintenance budget. The
BPN network estimated that PCI values were predominantly determined by the differences in
PCI, AADT and ESALs. Dynamic linear programming optimization estimated that CAD 150
million is the minimum annual budget to achieve most of the arterial and local roads in good
condition in Montreal.
Keywords: Pavement management; traffic simulation; backpropagation neural network;
performance modeling; uncertainties; linear programming; lifecycle optimization.
Introduction
The aging road network in Montreal City is at an advanced state of deterioration. Inappropriate
maintenance, low priority on infrastructure maintenance, and inadequate funding are the main
factors for this state of deterioration. Lack of comprehensive pavement management system
2
Page 3 of 35
(PMS) and a long-term plan are responsible for observing an increase in the budget for the City
of Montreal of more than 560% since 2001 (Amin and Amador-Jiménez 2014). The City needs a
holistic model of PMS to predict the response and performance of its pavements under predicted
dynamic traffic loads that optimizes the allocation of treatment operations.

Life-cycle cost analysis (LCCA) for PMS has been applied in a number of studies (Chan et al.
2008; Salem et al. 2003; Li and Madanu 2009). Chan et al. (2008) assessed LCCA practices in
the Michigan Department of Transportation and analyzed its accuracy in projecting the actual
costs and choosing the best alternatives for the treatments operations during the lifespan of the
pavement. Salem et al. (2003) applied risk-based life-cycle cost model that reflects the time to
failure of each treatment alternative and informs about the uncertainty levels that accompany the
estimated life-cycle costs. Li and Madanu (2009) developed an uncertainty-based method for
highway project-level LCCA to deal with computational uncertainty that inherited with input
factors. Zhang et al. (2013) applied life-cycle optimization model to develop a new network-
level pavement asset management system (PAMS) from historical values of pavement distress.
Traditional PMS tools such as PAVER and HDM3 are based on cost-benefit analysis that are
incapable of trading-off decisions between asset types and modes of transportation (Moazami et
al. 2011; Jain et al. 2005; Amin and Amador- Jiménez 2014; Humphries 2012). The linear
programming and other optimization techniques for PMS are capable of finding the optimal
solution of cost-effectiveness of maintenance and rehabilitation (M&R) operations and the
benefits of advancing or deferring a certain treatment (Hudson et al. 1997).
3
The Arizona Department of Transportation (ADOT) has applied Markovian chain–based
pavement management systems (PMS) to support its pavement design, construction, and
preservation activities (Li et al. 2006). Several other researchers also applied Markov decision
process (MDP) as a PMS tool (Abaza et al. 2004; Golabi and Pereira 2003; Gao and Zhang
2008). The MDP models optimize the M&R strategies based on the steady-state probabilities. In
reality, the pavements under a given maintenance policy usually takes many years to reach the
steady state and the proportion of the pavements are deteriorating or upgrading year-by-year
(Amin and Amador-Jiménez 2014).
Prediction of pavement deterioration, a crucial part of PMS, is explicitly complicated since the
pavement performance depends on a large number of dynamic and static attributes. Deterministic
and stochastic models are applied to predict the pavement deterioration. Deterministic models
are mechanistic, mechanistic-empirical and regression models (Archilla 2006; Yu et al. 2007;
Santos and Ferreira 2013; Sathaye et al. 2010). Stochastic models apply Markov transition
matrices that predict the ‘after’ condition of pavement knowing the ‘before’ condition (Ferreira
et al. 2011; Ortiz-García et al. 2006; Kobayashi et al. 2010). Deterministic and stochastic models
cannot properly address the uncertainties associated with data collection and computation. In
addition, MDP models suffer from somewhat unrealistic assumptions of discrete transition time
intervals and dependence of the future facility condition only on the current condition (Morcous
2002; Li et al. 2006). The MDP models lack the flexibility to consider the distinct conditions
associated with individual pavement projects or sections (Li et al. 2006).
4
Page 5 of 35
Ben-Akiva et al. (1993) developed the latent performance approach dealing with forecasting
uncertainties during condition data collection. This latent model had computational limitations
because number of outcomes, probabilities and computational effort to obtain M&R policies
increase exponentially with the number of distresses being measured. Attoh-Okine (1999)
proposed the Artificial Neural Network (ANN) for predicting the roughness progression in
flexible pavements. However, some built-in functions of ANN such as learning rate and
momentum term of ANN algorithm were not investigated properly. Several researchers also
applied ANN as a tool of PMS (Shekharan 1999; Yang et al. 2006). These models have not yet
overcome the functional limitations of neural network algorithms. This study accommodates the
simulated traffic and addresses the model uncertainties of pavement performance function.
This study develops the linear programming of PMS for the road network of Montreal City that
accommodates the simulated traffic during 50 years design period and deals with the uncertainty
of the pavement performance modeling.
Methodology
This study has been executed in three main steps: first the simulation of traffic (volumes and
truck loads), second the development of pavement performance models and third the
optimization of long term budget allocation and scheduling of interventions (Figure 1).
Figure 1: Flow chart of methodology
Simulation of traffic
Traffic volume, on each segment of the road network, has been simulated every five years for a
period of 50 years (between 2009 and 2058) through a travel demand model. A discrete choice
5
model estimates the trip generations from different Traffic Analysis Zones (TAZs) using
disaggregate household or individual level data. Trip generation is the function of gender, age,
personal and household income, occupation, family size, auto ownership, number of children in
the household, land use, and residential density. The model aggregates individual probabilities
(of individual trip making decisions) to predict the total number of trips produced in the TAZs
(Caliper 2005). The predicted trips are spatially distributed among TAZs by applying a doubly-
constrained gravity model (Equation 1) (Mozalin et al. 2000).
A multinomial Logit (MNL) model is applied to estimate the choice of modes by commuters,
assuming that the utility of an alternative is a function of the choice determinants, unknown
parameters and an i.i.d Gumbel-distribution error term. Choice determinants are travel time and
cost. Finally, a deterministic User Equilibrium (UE) model is applied to simulate the annual
average daily traffic (AADT) on each road segment of the network. The deterministic UE
method applies an iterative process to achieve a convergent solution so that no travelers can
improve their travel times (or cost) by shifting routes (Prashker and Bekhor 2004).
Pavement performance modeling
This study applies a Backpropagation Neural Network (BPN) method with a Generalized Delta
Rule (GDR) as a learning algorithm that minimizes uncertainty of pavements performance
modeling. The GDR learning algorithm is applied to the neural network because the relationship
is nonlinear and multidimensional (Freeman and Skapura 1991).The BPN network has two forms
of activation functions chosen from either the hyperbolic tangent or the logistic function. This
research uses the logistic function given that the output of the model (e.g. pavement condition) is
6
Page 7 of 35
a positive value (Freeman and Skapura 1991) and the logistic function results in positive outputs
in the unitary range (0, 1). Rather the hyperbolic tangent function in the range of (-1, 1). The
errors, estimated from the difference between calculated and observed outputs, are transferred
backward from the output layer in the ANN to each unit in the intermediate layer. Each unit in
the intermediate layer receives only a portion of the total error based roughly on the relative
contribution the unit made to the estimated output (Freeman and Skapura 1991). This process is
repeated layer-by-layer until each node in the network has received a proportion of error that
represents its relative contribution to the total error. The connection weights are updated based
on the error received by each unit (Freeman and Skapura 1991).
Pavement performance curves are estimated for four categories of roads based on the road
characteristics and the type of pavement: from now on arterial-flexible refers to an arterial road
with asphalt concrete pavement, in a similar fashion a local-flexible, arterial-rigid and local-rigid
refer to other combinations of either rigid or flexible pavements and road’s functional
classification. A Pavement Condition Index (PCI) was used in this research, in practical
applications it is advisable to replicate the work herein developed with independent indicators of
damage such as cracking and rutting, that can be easily correlated to available interventions, such
information was unfortunately not available.
For pavement deterioration, the input variables of the ANN for flexible pavements were: traffic
volume in the form of AADT and equivalent single axle loads (ESALs), structural number (SN),
pavement’s age (N) and difference of PCI between current and preceding year (∆PCI = PCI2009
– PCI2010). The ∆PCI helps to track the application of treatment operations during the preceding
7
year, and separate deterioration of intervened roads. Weather was constant across all road
segments and hence ignored on the case study, however, climate must be included when dealing
with roads traversing through various climate regions. The input variables for the rigid
pavements were AADT, ESALs, slab thickness (T), N and ∆PCI. Since AADT and ESALs are
log-linearly related to PCI, Log10 (AADT) and Log10 (ESALs) are taken as input variables of
PCI.
Data on pavement condition, age, thickness and road characteristics was collected from the City
of Montreal. From a total of 19826 segments in the network, the city counts with pavement
condition for 6868 segments in the year 2009 and 10842 segment in 2010. The AADT was
converted to 80-KN ESALs and only considers the total damage to the pavement caused by
commercial vehicles. The ESALs were calculated based on simulated AADT and traffic mixture
(number, type and distribution of commercial vehicles). Data on type and distribution of vehicles
on the road network of Montreal City are adopted from the report prepared by the Cement
Association of Canada (Cement Association of Canada 2012). The SN of the flexible pavements
is calculated from the thickness of pavement layers and climate condition of Montreal City
following the 1993 AASHTO Guide basic design equation.
Linear programming of pavement management system (PMS)
Lifecycle optimization to achieve and sustain acceptable mean network condition ( Q ) at a
minimum cost is used to find required levels of funding for Montreal road network (Equation 2
and 3). The maximization of total network condition under such a budget is then used to find
optimal strategic results for pavement management (Equation 4 and 5). This formulation relied
8
Page 9 of 35
on a decision tree containing all possible paths of pavement condition across time, after
hypothetically receiving available treatments (Amin and Amador 2014). This tree is based upon
a transfer function used to estimate condition (Qti) as a convex combination based on the decision
variable and the effectiveness or deterioration of the specific link on time t (Equation 6).
The objective function is to minimize cost (Z) and maximize pavement condition of the road
T a
network [ ∑∑ ( Li Qti ) ].
t =1 i =1
T a o
MINIMIZE Z = ∑∑∑ Ctj X tij Li (2)
t =1 i =1 j =1
T a a
Subject to : ∑∑ Li Qti ≥ (Q )∑ Li
t =1 i =1 i =1
(3)
T a
MAXIMIZE ∑∑ ( L Q ) i ti
t =1 i =1
(4)
T a k
Subject to: Z = ∑∑∑Ctj X tij Li ≤ Bt
t =1 i =1 j =1
(5)
0 ≤ Qt,i ≤ 100 and ∑X

j ∈ J ti
tij ≤ 1 {for all times t and for each asset i}
Qtij = Xtij (Q(t-1)ij + Eij)+ (1-Xtij) (Q(t-1)ij + Dit) (6)
Where Xtij is 1 if treatment j is applied on road segment i at year t, zero otherwise; Qti is condition
Index for road segment i at year t; Qtij is condition Index of road segment i at year t for treatment
j; Q(t-1)ij is condition Index of road segment i at year (t-1) for treatment j; Ctj is cost ($) of
treatment j at year t; Li is length of road (km) for road segment i; Eij is improvement (+) on road
9
segment i from treatment j, Dit is deterioration (-) on road segment i at time t, Bt is budget at year
t.
Data Analysis
Travel demand modeling and simulation of traffic volumes and loads
Traffic volume (AADT) is simulated for each road segment of Montreal road network at 5-years
interval during the fifty years analysis period by applying four-step travel demand model. Figure
2 shows the projected average AADT for four types of road categories during the analysis period.
It is estimated that the average traffic volume on arterial-flexible roads will increase from 8955
to 17069 during the period of 2009-2058 (Figure 2). The AADT will be increase to 4808, 11812
and 7684 on arterial-rigid, local-flexible and local-rigid roads during the same period,
respectively (Figure 2). Travel demand model at urban level is validated with original data of the
year 2008. Comparative evaluation of original and calibrated AADT shows that the overall
AADT on the road network of Montreal will be increased by only a 1.68% when applying the
travel demand model. In particular the model overestimates AADT for arterial-rigid and arterial-
flexible roads: 9.05% and 7.88% more, respectively. Conversely, the model underestimates
AADT for local-rigid and local-flexible roads: 5.13% and 4.40% less, respectively. The DUE
model assign traffic on a road segment based on the performance function (travel time) that is the
function of free-flow travel time, volume and capacity of road segment and calibration
parameters. This is why; increment of AADT is observed on the arterial roads applying the travel
demand model.
10
Page 11 of 35
The compounding effect from the traffic growth rate (1.01% annual growth) was removed from
predicted traffic in order to measure the accuracy of the model. The comparative analyses shown
that simulated traffic increases from 7.88% to 26.23% when compared to the AADT levels
calculated by the compound traffic growth rate on arterial-flexible roads during the period of
2009-2058. Figure 2 illustrates the other cases.

Figure 2: Simulated 50-percentile AADT for different road categories during the period of 2009-
2058
Accumulated traffic loads in the form of Equivalent Single Axles Loads (ESALs) were
calculated from predicted AADT and locally observed truck distributions combined with truck
factors (FHWA 2011). The model estimated 39.38, 11.56, 27.35 and 17.71 million ESALs for
arterial-flexible, arterial-rigid, local-flexible and local-rigid roads, for the period between 2009
and 2058 (Figure 3). During this period of 50 years, the simulated traffic explains otherwise non-
accounted increases of 26.80%, 18.32%, 13.59% and 6.80% of additional traffic loads (ESALs)
for the arterial-flexible, arterial-rigid, local-flexible and local-rigid roads (Figure 3). These
unaccounted differences could result in faster damage and inadequate timing of road
interventions.
Figure 3: Simulated 50-percentile ESALs (million) for different road categories during the period
of 2009-2058
Backpropagation Neural Network (BPN) for modeling of pavement performance
11
A Multi-Layer Perceptron (MLP) network was used to minimize the prediction error. The MLP
procedure computes the minimum and maximum values of the range and find the best number of
hidden layers to distribute the error term (IBM 2010). The MLP estimates the number of hidden
layers based on the minimum error in the testing data and the smallest Bayesian information
criterion (BIC) in the training data (IBM 2010). The logistic activation function is used for the
hidden layers so that the activation of the hidden unit is a Gaussian ‘bump’ as a function of input
units (IBM 2010).
The BPN network estimates that PCI values for arterial-flexible roads are predominantly
determined by ∆PCI and pavement’s age (Table 1). Other input variable such as log10 (AADT),
log10 (ESALs) and SN have 13.8 percent, 12 percent and 1.5 percent contributions in
determining the PCI value (Table 1). The ∆PCI also significantly influence the PCI values of
arterial-rigid, local-flexible and local-rigid roads by 33.1 percent, 33 percent and 32.9 percent
respectively (Table 1). However, pavement’s age does not significantly influence the PCI values
of arterial-rigid, local-flexible and local-rigid roads (Table 1).
[Table 1]
The log10 (AADT) and log10 (ESALs) have considerable importance to estimate the PCI values in
BPN models for arterial-rigid, local-flexible and local-rigid roads. For example, the log10
(AADT) has 23 percent importance to estimate PCI values of arterial-rigid with similar values
for the other categories (Table 1). The log10 (ESALs) variable contributes 24.8 percent of PCI
values for local-rigid roads (Table 1). The structural characteristics of pavement such as SN and
12
Page 13 of 35
slab thickness of flexible and rigid pavements do not have significant influence in determining
the PCI values respectively (Table 1). The reason is that the categorical values of thickness of
pavement’s layers for broader categories of AADT are applied in this study both for flexible and
rigid pavements from the report prepared by the Cement Association of Canada (2012). There is
a strong potential that the BPN models might estimate the significant or considerable influences
of SN and slab thickness on the PCI for flexible and rigid pavements respectively, if the actual
data on thickness of pavement’s layers for each road segment can be accommodated into the
BPN network.
The PCI value of each road segment is estimated based on the estimated relationship between
the PCI values and input variables applying BPN, and simulated traffic applying travel demand
modelling. The estimated PCI values are averaged for each road category for each period to
determine the average PCI value of arterial flexible, arterial rigid, local flexible and local rigid
roads. The average PCI values of the four road categories are plotted and fitted in the different
curves in the Figure 4 to show how the performance curves of the four road categories look like
during the 50 years analysis period. This study finds out that the polynomial and exponential
curves better represent the pattern of PCI curves for the selected four road categories.
Figure 4: Pavement performance curves for different road categories during the period of 2009-
2058
The original and estimated PCI values of Montreal road network for the year 2009 were
compared to validate the accuracy of the estimated PCI values using BPN model. The
13
comparison shows that the BPN models estimate on an average 3.26% less PCI values
compared to the original values. The lower estimation of PCI values is associated with the
higher estimation of AADT (1.68%) predicted by the travel demand model. The BPN models
estimate 3.23% and 4.48% lower PCI values comparing to the original PCI values for arterial-
rigid and arterial-flexible roads since these categories of roads observe 9.05% and 7.88%
increases of AADT, respectively. On the other hand, the BPN models estimate 3.74% and
1.61% lower PCI values comparing to the original PCI values for local-rigid and local-flexible
roads since these roads observe 5.13% and 4.40% decreases of AADT, respectively. Both the
original and estimated PCI values show that the overall structural quality of rigid pavements is
poor comparing to flexible pavements in Montreal Island.
Pavement treatment operations
Table 2 presents the criteria employed to identify the type of intervention that is applicable to
every segment at different points of time and following several possible paths (deteriorated or
improved). The criteria includes cost, effectiveness and range of applicability, these criteria came
from the Cement Association of Canada (2012). The timing of the intervention is modeled as a
binary decision variable (Equations 2 and 3) and the overall optimization algorithm takes care of
identifying the optimal set of interventions for the whole network during the analysis term (50
years in this case) within a complex structure with time dependencies that link the repercussions
of decisions through time. The aggregation of individual annual interventions found at the
solution return the minimum maintenance budget to achieve and sustain good pavement
condition. The computational time to find a global maximum was 48 seconds. It was found that
CAD 150 million is the minimum annual budget that ensures having good average roads
14
Page 15 of 35
condition (Figure 5). This research qualitatively defined pavement condition in four categories
such as excellent (PCI ≥80), good (80 > PCI ≥ 70), fair (70 > PCI ≥ 50) and poor (PCI<50).
Figure 6 shows that pavement condition will rapidly deteriorate after 31st year of the analysis
period under annual maintenance budget of CAD 125 million. Additional investment in
maintenance budget on the top of CAD 150 million will not significantly improve the pavement
condition rather the proportion of roads in good condition will be upgraded to excellent
condition. Figure 7 shows that the overall pavement condition of Montreal road network will not
be improved under the annual maintenance budget of CAD 175 million.
[Table 2]
Figure 5: Predicted conditions of roads after treatment operations under annual maintenance
budget of CAD 150 million
The predicted annual budget of CAD 150 million will almost equally be distributed for the
treatment operations of flexible and rigid pavements during first 20 years, but flexible pavements
will require more maintenance budget during the period of 2029 -2045 (Figure 8). Considerably
higher maintenance budget will be invested for treating rigid pavements after 36th design period
(Figure 8). The majority of the maintenance budget for flexible pavements will be invested in
15
reconstruction (RC), resurfacing (RS), repair and overlay during the first 6 years of the analysis
period. For the remaining of the analysis period, annual budget for flexible pavements will be
invested in rout and sealing (CS) treatment (Figure 9). Similar pattern of budget distribution is
found in treatment operations for rigid pavements (Figure 10).

Figure 8: Distribution of annual maintenance budget (CAD $150 million) among rigid and
flexible pavements
Figure 9: Distribution of annual maintenance budget for different treatment operations of flexible
pavements
Figure 10: Distribution of annual maintenance budget for different treatment operations of rigid
pavements
It is worthwhile mentioning that the herein proposed model improves over the conventional PMS.
Firstly, it bases performance on dynamic traffic. Traditional methods use deterioration curves
that are based on the historical traffic (volume and growth). But traffic volume and distribution is
related to land use, economy, employment opportunities, and travel behavior. This study predicts
dynamic traffic volume and loads by applying travel demand modeling. Secondly, this proposed
model deals with the uncertainties of developing pavement performance curves. Traditional
deterministic and stochastic methods for predicting pavement deterioration are not only unable to
address dynamic traffic loads, but also have drawbacks of model uncertainty, subjective
evaluations of pavement condition and steady-state probabilities for transition of pavement
condition from one state to another. This model helps the transportation authorities to manage
continuous aggregate behavior of the transportation system, estimate more accurate pavement
16
Page 17 of 35
deterioration and solve lifecycle optimization problems of pavement management at any time
interval during the lifespan of pavement.
Conclusion
Traditional PMS methods are limited to a static traffic prediction and inaccurate deterioration
models, erroneously biasing investment decisions of pavement interventions. This study
develops a linear programming approach for PMS that accommodates simulated traffic during a
long term analysis period. Even if dynamic traffic is considered, PMS models still do not address
the uncertainty of the predictions. This research proposes a method to address uncertainty on
pavement deterioration predictions. A case study of Montreal road network is used to illustrate
the proposed method. Arterial roads of Montreal City, mostly constructed in 1950’s, were found
at an advanced state of deterioration. Linear programming was applied to develop M&R
strategies ensuring the good pavement condition of roads at a minimum maintenance budget.
Lifecycle optimization of PMS estimated that CAD 150 million is the minimum annual budget to
achieve most of arterial and local roads are at least in good condition in Montreal City. The
developed model of PMS has two-fold improvement on the conventional methods of PMS.
Firstly, this study predicts dynamic traffic volumes and loads during the fifty-year analysis
period by applying travel demand models. Secondly, this model deals with the computational
errors of developing the pavement performance curves. This proposed model will help the
transportation authorities to manage continuous aggregate behavior of transportation system,
estimate more accurate pavement deterioration and solve lifecycle optimization problems of
pavement management at any time interval during the lifespan of pavement.
17
The fast rate of deterioration of Montreal roads is confirmed by the estimations from the model:
even though overall vehicle traffic is expected to double within 50 years, truck traffic is expected
to suffer a much faster increase; doubling the number of ESALs every 15 years, this is a direct
result from economic activity and some traffic could be regional. A backpropagation Neural
Network (BPN) of the type multilayer perceptron combined with a Generalized Delta Rule
(GDR) was able to improve the estimation of future deterioration of pavements. Both rigid and
flexible pavements deteriorated at a similar rate, no big differences were observed. A budget of
at least CAN$150 million is required to sustain arterial and local roads of Montreal in good
condition.
Roads in the island of Montreal need to undergo through a stabilization period for about 25
years, a steady state seems to be reached after that and only preventive maintenance treatments
are applied after that. Future research should study more complex maintenance rules regarding
the limited back-to-back application of certain interventions, per instance limiting the number of
consecutive crack sealing before enforcing an overlay. Future research can study alternate
methods for traffic prediction. Future research should use individual distress indicators of
pavement damage instead of PCI.
References
Abaza, K., Ashur, S., and Al-Khatib, I. 2004. Integrated Pavement Management System with a
Markovian Prediction Model. Journal of Transportation Engineering, 130(1), 24–33.
18
Page 19 of 35
Amin, M.S.R., and Amador, L.E. 2014. The Multi-criteria based Pavement Management System
for Regional Road Network of Atlantic Provinces of Canada. International Journal of
Pavements, 13(1-2-3). [In Press]
Amin, M.S.R., and Amador-Jiménez, L. 2014. A performance-based Pavement Management
System for the road network of Montreal city—a conceptual framework. In Asphalt
Pavements. Edited by Y.R. Taylor & Francis Group, London, pp. 233-244.
Archilla, A.E. 2006. Repeated Measurement Data Analysis in Pavement Deterioration Modeling.
Journal of Infrastructure Systems, 12(3):163–173.
Attoh-Okine, N.O. 1999. Analysis of Learning Rate and Momentum Term in Backpropagation
Neural network Algorithm Trained to Predict Pavement Performance. Advances in
Engineering Software, 30:291-302.
Ben-Akiva, M., Humplick, F., and Madanat, S. 1993. Infrastructure Management under
Uncertainty: The Latent Performance Approach. Journal of Transportation Engineering,
119(1): 43-58.
Cement Association of Canada. 2012. Methodology for the Development of Equivalent
Pavement Structural Design Matrix for Municipal Roadways- Montréal and Québec City:
Including Maintenance & Rehabilitation Schedules and Life Cycle Cost Analysis. Final
Report. Applied Research Associates Inc., Ottawa, ON.
Chan, A., Keoleian, G. and Gabler, E. 2008. Evaluation of lifecycle cost analysis practices used
by the Michigan Department of Transportation. Journal of Transportation Engineering, 134
(6): 236–245.
Freeman, J.A., and Skapura, D.M. 1991. Neural Networks. Algorithms, Applications, and
Programming Techniques. Addison-Wesley Publishing Company, Reading, MA.
19
Federal Highway Administration. 2011. FHWA vehicle types. Available from
http://www.fhwa.dot.gov/policy/ohpi/vehclass.htm (accessed 24 April 2012).
Hudson, W.R., Haas, R., and Uddin, W. 1997. Infrastructure Management. McGraw Hill, New
York, N.Y.
GAO, L., and Zhang, Z. 2008. Robust Optimization for Managing Pavement Maintenance and
Rehabilitation. Transportation Research Record: Journal of the Transportation Research
Board, 2084: 55–61.
Golabi, K. and Pereira, P. 2003. Innovative Pavement Management and Planning System for
Road Network of Portugal. Journal of Infrastructure Systems, 9(2): 75–80.
Humphries, C. 2012. Examination of potential cost benefits for converting surfaced roads to
unsurfaced roads in Texas. M.Sc. thesis, Dept. of Engineering Technology, Texas State
University-San Marcos.
IBM 2010. IBM SPSS Neural Networks 19. Available from
http://www.sussex.ac.uk/its/pdfs/SPSS_Neural_Network_19.pdf (accessed 10 January 2012).
Jain, S.S., Aggarwal, S., and Parida, M. 2005. HDM-4 Pavement Deterioration Models for Indian
National Highway Network. Journal of Transportation Engineering, 131(8):623–631.
Kobayashi, K., Do, M., and Han, D. 2010. Estimation of Markovian transition probabilities for
pavement deterioration forecasting. KSCE Journal of Civil Engineering, 14(3): 343-351.
Li, Z. and Madanu, S. 2009. Highway project level life-cycle benefit/cost analysis under
certainty, risk, and uncertainty: methodology with case study. Journal of Transportation
Engineering, 135(8): 516–526.
20
Page 21 of 35
Li, Y., Cheetham, A., Zaghloul, S., Helali, K., and Bekheet, W. 2006. Enhancement of Arizona
Pavement Management System for Construction and Maintenance Activities. Transportation
Research Record: Journal of the Transportation Research Board, 1974:26–36.
Moazami, D., Behbahani, H., and Muniandy, R. 2011. Pavement rehabilitation and maintenance
prioritization of urban roads using fuzzy logic. Expert Systems with Applications, 38:12869–
12879.
Morcous, G., Rivard, H., and Hanna, A.M. 2002. Modeling bridge deterioration using case-based
reasoning. Journal of Infrastructure Systems, 8(3):86–95.
Mozalin, M., Thill, J.-C., and Usery, E.L. 2000. Trip distribution forecasting with multilayer
perceptron neural networks: A critical evaluation. Transportation Research Part B:
Methodological, 34(1): 53-73.
Ortiz-García, J.J., Costello, S.B., and Snaith, M.S. 2007. Derivation of Transition Probability
Matrices for Pavement Deterioration Modeling. Journal of Transportation Engineering,
132(2):141-161.
Prashker, J.N., and Bekhor, S. 2004. Route Choice Models Used in the Stochastic User
Equilibrium Problem: A Review. Transport Reviews, 24(4):437-463.
Salem, O., Abourizk, S., and Ariaratnam, S. 2003. Risk-based life-cycle costing of infrastructure
rehabilitation and construction alternatives. Journal of Infrastructure Systems, 9(1):6–15.
Santos, J., and Ferreira, A. 2013. Life-cycle cost analysis system for pavement management at
project level. International Journal of Pavement Engineering, 14(1):71–84.
Sathaye, N., Horvath, A., and Madanat, S. 2010. Unintended impacts of increased truck loads on
pavement supply-chain emissions. Transportation Research Part A, 44:1–15.
21
Shekharan, A.R. 1999). Assessment of Relative Contribution of Input Variables to Pavement
Performance Prediction by Artificial Neural Networks. Transportation Research Record:
Journal of the Transportation Research Board, 1655:35-41.
Yang, J., Gunaratne, M., Lu, J.J., and Dietrich, B. 2007. Use of Recurrent Markov Chains for
Modeling the Crack Performance of Flexible Pavements. Journal of Transportation

Engineering, 133(6):347–354.
Yang, J., Lu, J.J., Gunaratne, M., and Dietrich, B. 2006. Modeling Crack Deterioration of
Flexible Pavements. Comparison of Recurrent Markov Chains and Artificial Neural
Networks. Transportation Research Record: Journal of the Transportation Research Board,
1974:18–25.
Yu, J., Chou, E.Y.J., and Luo, Z. 2007. Development of Linear Mixed Effects Models for
Predicting Individual Pavement Conditions. Journal of Transportation Engineering,
133(6):347–354.
Zhang, H., Keoleian, G.A., and Lepech, M.D. 2013. Network-Level Pavement Asset
Management System Integrated with Life-Cycle Analysis and Life-Cycle Optimization.
Journal of Infrastructure Systems, 19(1):99–107.
22
Page 23 of 35
Data Input
Travel demand modelling • Disaggregate household data
on demographic and socio-
Simulated • Trip generation – Discrete
economic characteristics of
Annual choice model commuter
Average Daily • Trip distribution - Doubly- • Auto ownership
Traffic (AADT) constrained gravity model • Land use and residential
• Mode choice - Multinomial density
Logit (MNL) model • Total number of travelers in
• Traffic assignment – different income groups
Deterministic User • Travel time and cost
Data Input
• Simulated AADT
• Distribution of commercial vehicles
Traffic Loads Equivalent Single Axle • Load Equivalency Factor (LEF)
loads (ESALs) • Directional split and lane distribution factor
• Number of Commercial Trucking Days per Year
• Design periods – 50 years
Flexible pavement
Data Input
• Simulated AADT and
ESALs
Pavement
• Structural number
Performance
(SN)
Modeling
(Backpropagation • Pavement’s age (N)
Neural Network) • Slab thickness (T)
• Difference of PCI
between current and
preceding year, ∆PCI
Rigid pavement
Data Input
• Predicted PCI
• Treatment operations
• Unit cost of treatment
operations
Pavement
management Linear programing of life-cycle • Length of road segments
system optimization • Annual maintenance
budget
al use only. This Just-IN manuscript is the accepted manuscript prior to copy editing and page composition. It may differ from the final official version
AADT
0
2000
4000
6000
8000
10000
12000
14000
16000
18000
2008
2013
Local-rigid
Arterial-rigid
2018
Arterial-flexible
2023
2028
Local-Flexible
Arterial-flexible
Years
2033
2038
Local-rigid Traffic growth rate

2043
Arterial-rigid
Local-Flexible
2048
2053
2058
Page 24 of 35
Page 25 of 35
Cumulative ESALs (million)
0
5
10
15
20
25
30
35
40
2008
2013
2018
Local-Flexible
Arterial-flexible
2023
2028
Local-Flexible
Arterial-flexible
2033
Years
2038
Local-rigid
Arterial-rigid
2043
Local-rigid
2048
Arterial-rigid
2053
2058
al use only. This Just-IN manuscript is the accepted manuscript prior to copy editing and page composition. It may differ from the final official version Page 26 of 35
77
75
74
70 70
65
R² = 0.9937
60
67
55
Arterial Flexible
R² = 0.9921 Arterial Rigid

50
Local Flexible
R² = 0.9642 Local Rigid
PCI
45
Expon. (Arterial Flexible)
R² = 0.9644 Poly. (Arterial Rigid)
40
Expon. (Local Flexible)
Expon. (Local Rigid)
35 30
30
28
25 27
20 25
2008 2013 2018 2023 2028 2033 2038 2043 2048 2053 2058
Years
Page 27 of 35
Length (meters)
Periods
Length (meters)
Periods
Page 28 of 35
Page 29 of 35
Length (meters)
Periods
Maintenance budget (CAD)
Total cost
Periods
Total costs for rigid pavements
Total cost for flexible pavements
Page 30 of 35
Page 31 of 35
(Local)
(Arterial)
Mill HMA (Local)

Mill HMA (Arterial)
Rout and seal (Local)
Rout and seal (Arterial)
Resurface with ESG 10 (Local)

Resurface with ESG 10 (Arterial)
Spot repairs, mill 40 mm/ patch 40 mm
Spot repairs, mill 40 mm/ patch 40 mm
Periods
Periods
Reseal joints (Local)
Reseal joints (Arterial)
Reconstruction (Local)
Reconstruction (Arterial)
Full depth PCC repair (Local)
Full depth PCC repair (Arterial)
Partial depth PCC repair (Local)
Partial depth PCC repair (Arterial)
Page 32 of 35
Page 33 of 35
Table 1 Importance of input variables to estimate PCI values in BPN networks

Arterial Collector
Input variables
Flexible Rigid Flexible Rigid
∆PCI .364 .331 .330 .329
Log10(AADT) .138 .230 .226 .201
Log10(ESALs) .120 .194 .221 .248
Pavement’s Age (N) .363 .162 .123 .211
Structural Number (SN) .015 .100
Slab Thickness (mm), T .083 .012

al use only. This Just-IN manuscript is the accepted manuscript prior to copy editing and page composition. It may differ from the final official version Page 34 of 35
Table 2: Treatment and Operational Windows Used in Network-Level Trade-Off Analysis

Pavement Unit cost (CAD$)
Type Treatments Operational window Arterial Local
AGE ≤ 5
Reseal joints, % length (m) 80 ≥ PCI (Arterial) ≥ 77; 80 ≥ PCI (Local) ≥ 73 10 10
5 ≤ AGE ≤ 12
Partial depth PCC repair, % area, (sq. m.) 76 ≥ PCI (Arterial) ≥ 68; 72 ≥ PCI (Local) ≥ 55 150 150
12 ≤ AGE ≤ 25
Full depth PCC repair, % area, (sq. m.) 67 ≥ PCI (Arterial) ≥ 56; 54 ≥ PCI (Local) ≥ 44 125 125
Reconstruction - Arterial 108
Rigid 200 mm PCC pavement, 25.4mm dowels (m²) 64
Base - MG 20, mm (t) 23
Subbase - MG 112, mm (t) AGE ≥ 26; PCI (Arterial) < 55 21
Reconstruction - Local 98.75
175 mm PCC pavement, no dowels (m²) 54.75
Base - MG 20, mm (t) AGE ≥ 26; PCI (Local) < 43 23
Subbase - MG 112, mm (t) 21
AGE ≤ 5
Rout and seal, m/km (m) 80 ≥ PCI (Arterial) ≥ 76; 80 ≥ PCI (Local) ≥ 72 5 5
Spot repairs, mill 40 mm/ patch 40 mm, % area 5 ≤ AGE ≤ 10
(sq. m.) 74 ≥ PCI (Arterial) ≥ 69; 70 ≥ PCI (Local) ≥ 65 20 20
10 ≤ AGE ≤ 15
Mill HMA, mm (t) 68 ≥ PCI (Arterial) ≥ 64; 64 ≥ PCI (Local) ≥ 59 10.4 10.4
15 ≤ AGE ≤ 25
Resurface with ESG 10, mm (t) 63 ≥ PCI (Arterial) ≥ 58; 58 ≥ PCI (Local) ≥ 46 135 135
Reconstruction - Arterial 179
Flexible HMA - ESG 10, mm (t) 70-28 135
Base - MG 20, mm (t) 23
Subbase - MG 112, mm (t) AGE ≥ 26; PCI (Arterial) < 58 21
Reconstruction - Local 173
HMA - ESG 10, mm (t) 64-28 129
Base - MG 20, mm (t) 23
Subbase - MG 112, mm (t) AGE ≥ 26; PCI (Local) < 46 21
Page 35 of 35
List of figures
Figure 1: Flow chart of methodology
Figure 2: Simulated 50-percentile AADT for different road categories during the period of 2009-
2058
Figure 3: Simulated 50-percentile ESALs (million) for different road categories during the period
of 2009-2058
Figure 4: Pavement performance curves for different road categories during the period of 2009-
2058
Figure 8: Distribution of annual maintenance budget (CAD $150 million) among rigid and
flexible pavements
Figure 9: Distribution of annual maintenance budget for different treatment operations of flexible
pavements
Figure 10: Distribution of annual maintenance budget for different treatment operations of rigid
pavements

Pavement Management With Dynamic Traffic and ANN: A Case Study of Montreal

Uploaded by

Document Information

Original Description:

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Pavement Management With Dynamic Traffic and ANN: A Case Study of Montreal

Uploaded by

Copyright:

Available Formats

Page 1 of 35

Pavement Management with dynamic traffic and ANN: a case study of

Md. Shohel Reza Amin

Research Assistant, Dept. of Building, Civil and Environmental Engineering,

Concordia University, 1515 St. Catherine Ouest

Montreal, QC H3G1M8, Canada

Tel: +14389364119. Fax: +1-514-848-7965.

Associate Professor, Dept. of Building, Civil and Environmental Engineering,

Concordia University, 1515 St. Catherine Ouest

Montreal, QC H3G1M8, Canada

Tel: +1-514-848-2424 (ext. 5783). Fax: +1-514-848-7965.

Pavement Management with dynamic traffic and ANN: a case study of

This study improves pavement management system developing a linear programming

develop pavement performance models without uncertainties. Linear programming of lifecycle

optimization is applied to develop maintenance and rehabilitation strategies to ensure the

Keywords: Pavement management; traffic simulation; backpropagation neural network;

performance modeling; uncertainties; linear programming; lifecycle optimization.

dynamic traffic loads that optimizes the allocation of treatment operations.

solution of cost-effectiveness of maintenance and rehabilitation (M&R) operations and the

benefits of advancing or deferring a certain treatment (Hudson et al. 1997).

The Arizona Department of Transportation (ADOT) has applied Markovian chain–based

(Amin and Amador-Jiménez 2014).

associated with individual pavement projects or sections (Li et al. 2006).

of the pavement performance modeling.

Figure 1: Flow chart of methodology

constrained gravity model (Equation 1) (Mozalin et al. 2000).

Pavement performance modeling

Rule (GDR) as a learning algorithm that minimizes uncertainty of pavements performance

on the error received by each unit (Freeman and Skapura 1991).

information was unfortunately not available.

following the 1993 AASHTO Guide basic design equation.

Linear programming of pavement management system (PMS)

Lifecycle optimization to achieve and sustain acceptable mean network condition ( Q ) at a

0 ≤ Qt,i ≤ 100 and ∑X

Qtij = Xtij (Q(t-1)ij + Eij)+ (1-Xtij) (Q(t-1)ij + Dit) (6)

Travel demand modeling and simulation of traffic volumes and loads

2009-2058. Figure 2 illustrates the other cases.

Backpropagation Neural Network (BPN) for modeling of pavement performance

units (IBM 2010).

of arterial-rigid, local-flexible and local-rigid roads (Table 1).

poor comparing to flexible pavements in Montreal Island.

Pavement treatment operations

be improved under the annual maintenance budget of CAD 175 million.

budget of CAD 150 million

budget of CAD 125 million

budget of CAD 175 million

found in treatment operations for rigid pavements (Figure 10).

evaluations of pavement condition and steady-state probabilities for transition of pavement

interval during the lifespan of pavement.

models, erroneously biasing investment decisions of pavement interventions. This study

at an advanced state of deterioration. Linear programming was applied to develop M&R

transportation authorities to manage continuous aggregate behavior of transportation system,

pavement management at any time interval during the lifespan of pavement.

pavement damage instead of PCI.

Markovian Prediction Model. Journal of Transportation Engineering, 130(1), 24–33.

for Regional Road Network of Atlantic Provinces of Canada. International Journal of

Pavements, 13(1-2-3). [In Press]

Amin, M.S.R., and Amador-Jiménez, L. 2014. A performance-based Pavement Management

Journal of Infrastructure Systems, 12(3):163–173.

Neural network Algorithm Trained to Predict Pavement Performance. Advances in

Engineering Software, 30:291-302.

Uncertainty: The Latent Performance Approach. Journal of Transportation Engineering,